Developing a simple field triage scheme compatible with regional trauma situations and medical systems using machine learning: A nationwide trauma database study

Abstract Background and Objectives Trauma patients' characteristics and prehospital medical systems differ between countries. We aimed to establish a simple field triage scheme for identifying severe injury patients that reflect the regional features of trauma demographics and trauma management systems. Methods This was a retrospective cohort study using the Japan Trauma Data Bank (JTDB). The main outcome was all‐cause in‐hospital mortality. Blunt trauma patients ≥18 years of age were enrolled. Thirty‐five parameters (including demographic characteristics, mechanism of injury, vital signs, and abbreviated injury scale 98) were assessed for each patient using a machine learning method. After calculating the feature importance of each variable from the JTDB 2019–2021 dataset, we selected variables and made schemes by including different combinations of variables. The schemes were analyzed using a receiver operating characteristic curve and validated using the JTDB 2004–2018 dataset. Results We analyzed 48,194 subjects in JTDB 2019–2021. Eight variables were selected and schemes calculated and evaluated using 185,728 cases in the JTDB 2004–2018 dataset. The area under the curve of the most accurate and simplest scheme was 0.845. When the cut‐off value was set at 6, the mortality rate of the positive group (i.e., score ≥6) was 31.8%, the undertriage rate was 4.6%, and the overtriage rate was 68.2%. Conclusions We established an easy and reliable field triage scheme suited to the Japanese trauma system which comprised disturbance of consciousness, respiratory failure, shock, head injury, chest injury, multiple injury, fall from height, and age.